Can AI Actually Generate Patentable Inventions?

Yes — with one structural requirement. An AI system cannot be named as an inventor, but inventions developed with AI assistance are patentable when a human makes a significant contribution to their conception. Structure the human role correctly and AI becomes the most productive invention instrument ever built: the AIMS program has 78 USPTO-filed provisional patents and roughly 1,275 claims to show for it.

This is the question every founder asks before taking AI research seriously, and it deserves a straight answer rather than hedging. Here's the legal landscape, what patentability actually requires, and how the AIMS methodology structures its process so the output is ownable.

What does the law actually say?

Two facts anchor everything:

• AI cannot be an inventor. In Thaler v. Vidal (2022), the U.S. Court of Appeals for the Federal Circuit held that a patent inventor must be a natural person. Patent offices in most major jurisdictions reached the same conclusion. Naming an AI system as inventor gets an application rejected outright.
• AI-assisted inventions are patentable. USPTO guidance issued in 2024 confirmed that using AI in the inventive process does not disqualify a patent — as long as a natural person made a significant contribution to the invention. The analysis runs claim by claim: for each claim, some human must have contributed significantly to its conception.

So the question isn't "can AI invent?" in the abstract. It's "is your process structured so that a human is genuinely doing the inventing, with AI as the instrument?" That's a process-design question, and it's answerable.

What makes any invention patentable?

The classic three-part test, unchanged by AI: the invention must be novel (not already disclosed in prior art), non-obvious (not a trivial step for a person skilled in the field), and useful. AI changes none of the criteria — what it changes is your ability to verify them before filing. An agent that has exhaustively mapped the literature and existing claims can tell you where the prior-art boundaries actually sit, which turns novelty from a lawyer's guess into a mapped gap. That mapping step is baked into the AIMS pipeline as the IP-gap analysis in stage 5.

How does AIMS keep the human contribution significant?

Deliberately, at four points in the pipeline — and this design is what makes portfolio-scale filing defensible:

• Problem conception. A human defines the research question and the constraint set — what the material must do, at what cost, under which conditions. Framing the problem is inventive work, and it happens before any model runs.
• Kill-or-proceed decisions. Through premise verification and structured falsification, a human decides which candidates die and which advance. Judgment under uncertainty is exactly the contribution inventorship doctrine cares about.
• Candidate selection and claim strategy. Which surviving candidates become claims, how broad the claims run, and what gets filed versus held — human calls, documented.
• Adversarial review. Stage 8 invites domain-expert falsification of every paper before publication. Surviving structured criticism is part of what makes the eventual claims defensible rather than decorative.

The pattern: AI searches, drafts, and maps at superhuman scale; humans conceive, judge, and decide. That division isn't a legal workaround — it's also just how good research works.

What does the proof look like?

Receipts, not theory. Running this process across 40 scientific domains, AIMS has filed 78 provisional patents carrying roughly 1,275 claims — covering PFAS water purification membranes, iron-air battery electrodes, lead-free radiation shielding, topological quantum computing architectures, non-toxic fire retardants, and dozens more. The methodology itself is patented (Patent 14). Forty-one research papers document the work, with published papers carrying permanent DOIs on Zenodo.

Note what those filings are: provisional applications — priority-date protection on novel claims, filed before publication, with twelve months to convert the strongest ones. That's the deliberate strategy of a portfolio builder, and the full workflow is laid out in how to build an IP portfolio with AI research agents.

What are the honest caveats?

Three, because overclaiming poisons credibility:

• A filed provisional is not a granted patent. Provisionals aren't examined; their value is priority and optionality. Claims prove out through conversion, prosecution, and — ultimately — licensing or litigation.
• Computational claims still want experimental validation. AIMS research is computational with clear experimental pathways; a claim backed by a confirmed experiment is stronger than one backed by screening alone. This is precisely why AIMS seeks institutional lab partners.
• Unfalsified AI output is a liability, not an asset. Filing on a hallucinated result wastes money and credibility. The falsification stage isn't optional — see how to falsify AI-generated scientific claims.

For the broader context on how the discovery side works, start with what AI-driven materials discovery is.

FAQ

Can an AI system be named as an inventor on a patent?

No. U.S. courts held in Thaler v. Vidal (2022) that an inventor must be a natural person, and patent offices in most major jurisdictions have taken the same position. AI cannot be an inventor — but that does not make AI-assisted inventions unpatentable.

Are AI-assisted inventions patentable in the United States?

Yes. USPTO guidance issued in 2024 confirms that inventions developed with AI assistance are patentable when a natural person made a significant contribution to the invention. The analysis is claim-by-claim: for each claim, a human must have contributed significantly to its conception.

What proof exists that AI-assisted research produces filed patents?

The AIMS program is a working example: 78 USPTO-filed provisional patents carrying roughly 1,275 claims across 40 scientific domains, all produced through an AI-driven constraint-satisfaction methodology with human-owned constraint definition, falsification decisions, and claim strategy.

How do you keep AI-assisted inventions defensible?

Structure the human contribution deliberately: humans define the problem constraints, make the kill-or-proceed decisions during falsification, select which candidates become claims, and shape the claim language. Document that process. The AI is a search and drafting instrument; conception decisions stay human.